EditSSC

Official code release for our CVPR Workshop paper.

Project Page

• Webpage

Setup

Conda (recommended)

cd /path/to/EditSSC
conda env create -f environment.yml
conda activate editssc

Autoencoder Training

Run:

python script/train_ae_main.py --config configs/common_ae_base.yaml

Save Triplanes

Run triplane extraction:

python script/save_triplane.py --config configs/common_ae_base.yaml

Diffusion Training

Train an unconditional diffusion model on the saved triplanes:

python script/train_diffusion_main.py --config configs/common_diffusion_base.yaml

Train a LiDAR-conditioned diffusion model :

python script/train_diffusion_main.py --config configs/common_diffusion_cond_lidar.yaml

Sample Generation

Generate scenes from a trained diffusion checkpoint.

python generation/generate_samples.py --config configs/common_diffusion_base.yaml

Sketch guided generation

Training-free scene generation from a BEV sketch (canvas). The pipeline has two steps: first build a class ↔ VQ-VAE code mapping, then run inpainting with a sketch layout.

Step 1 — Analyze VQ-VAE code ↔ class correspondences

Run this on your trained autoencoder. It measures which VQ-VAE codebook entries correspond to each semantic class and writes two files :

• vqvae_codes_analysis.txt — human-readable report
• vqvae_codes_99_coverage.json — mapping used by sketch generation (~99% class coverage)

python generation/analyze_vqvae_codes.py --config configs/common_ae_base.yaml

Step 2 — Generate scenes from a sketch (training_free_gen.py)

Pass the JSON from step 1 to map each canvas pixel (semantic class id) to a VQ-VAE code, then inpaint with the diffusion model.

List built-in canvas layouts:

python generation/training_free_gen.py --list

Run with one or several predefined layouts (roundabout, S-road, cross-road, etc.):

python generation/training_free_gen.py \
  --config configs/common_diffusion_base.yaml \
  --codes-json models/semantic_ae/common_ae_base/vqvae_codes_99_coverage.json \
  --canvas roundabout

Custom canvas

You can also use your own BEV sketches by adding a builder in generation/training_free_gen.py:

1. Define a function that returns a (128, 128) NumPy array of SemanticKITTI train ids (e.g. 0 empty, 9 road, 15 vegetation, 1 car — see canvas builders in the same file).
2. Register it in CANVAS_REGISTRY with a build_fn and save_dir_suffix.
3. Run with --canvas your_layout_name as above.